Title: Application-Driven Solutions for Exploiting Co-processors in High Performance Computing

Sara Karamati
Computer Science PhD Student
School of Computational Science and Engineering
Georgia Institute of Technology

Date: Wednesday, April 10th, 2024
Time: 12:00pm – 2:00pm (EDT)
Meeting Link: Zoom Link


Committee:
Dr. Richard W. Vuduc (advisor), School of Computational Science and Engineering, Georgia Institute of Technology
Dr. Jeffrey Young, School of Computer Science, Georgia Institute of Technology
Dr. Hyesoon Kim, School of Computer Science, Georgia Institute of Technology
Dr. Spencer Bryngelson, School of Computational Science and Engineering, Georgia Institute of Technology
Dr. Chris Siefert, Sandia National Laboratories

Abstract:
In this proposal, we demonstrate novel strategies and insights for optimizing high-performance applications, addressing algorithmic imbalances by tailoring to the unique capabilities of modern co-processors like Graphics Processing Units (GPUs) or Data Processing Units (DPUs). We investigate the unexpected potential of DPUs as compute accelerators in HPC applications far beyond their conventional roles in networking control, storage management, or security.

Our exploration is structured around case studies exemplifying distinct algorithmic complexities: Single Source Shortest Path (SSSP) in graph computing, MiniMD in molecular dynamics simulation, and Maxwell solver in electromagnetics applications. Each study presents challenges in task parallelism, workload distribution, and data dependencies, necessitating precise algorithmic modifications to counteract imbalances and inefficiencies. 

This work unfolds in three pivotal sections, each dedicated to one of the aforementioned case studies. Initially, we delve into optimizing a GPU-based Single-Source Shortest Path (SSSP) algorithm by modulating an algorithmic parameter, delta, for fine-tuning parallelism and power usage, thereby realizing up to 50% speedup and 25% power savings. Secondly, we investigate the optimization of the MiniMD Molecular Dynamics Simulation, proposing a new heuristic devised to enhance task concurrency and a new offloading scenario to DPUs. Despite the DPUs' typically modest performance, we attain up to 20% performance boost, albeit with a moderate increase in power cost. Finally, we propose a task partitioning strategy for the Maxwell Multigrid Solver, accentuating the practical and theoretical implications of task offloading and partitioning on DPUs, paving the way for an informed, nuanced approach to enhancing computational performance in heterogeneous environments. Finally, through the MiniMD and Maxwell solver studies, we investigate the efficiency and performance of DPUs, leading to a careful examination of the versatility of DPUs across different computational scenarios.